{
  "schema_version": "1.0",
  "id": "retrieval-index-structures-2026-10-28",
  "title": "Set postings versus sorted compact arrays",
  "published_at": "2026-10-28",
  "evidence_status": "reproducible synthetic benchmark",
  "workload": "Synthetic document collection at 2,000, 8,000 and 20,000 documents; eight known-label queries; three repetitions per configuration.",
  "results_20000_documents": {
    "set_postings": {
      "median_build_seconds": 2.686,
      "median_query_ms": 6.065,
      "python_tracemalloc_peak_mib": 106.53
    },
    "sorted_compact_arrays": {
      "median_build_seconds": 2.129,
      "median_query_ms": 1.864,
      "python_tracemalloc_peak_mib": 12.81
    }
  },
  "ranking_agreement": "Identical rankings on all eight fixture queries in the tested configurations.",
  "limitations": [
    "Synthetic fixture and deliberately easy known-label queries; hit@10 is a mechanics check, not real-world relevance.",
    "tracemalloc peak is not process RSS.",
    "Single-host pilot; production defaults should not be changed from this result alone."
  ],
  "artifacts": {
    "report": "https://darkoza.com/science/reports/retrieval-index-structures/report.md",
    "csv": "https://darkoza.com/science/reports/retrieval-index-structures/results.csv",
    "raw_results": "https://darkoza.com/science/reports/retrieval-index-structures/results.json",
    "source": "https://darkoza.com/science/reports/retrieval-index-structures/benchmark.py"
  }
}